Odds ratios of treatment response were well approximated from continuous rating scale scores for meta-analysis

To empirically evaluate the performance of methods for estimating odds ratios and their corresponding standard errors from continuous end point data for meta-analysis.

A database of randomized controlled trials of chronic depression treatments was used. Trials that reported both continuous and dichotomous end points for symptom improvement were considered. Odds ratios and standard errors were calculated from the dichotomous data and estimated from the continuous data using currently available methods: Hasselblad and Hedges (HH), Cox and Snell (CS), Furukawa (F), Suissa (S), and Kraemer and Kupfer (KK). Single and meta-analytically pooled observed and estimated values were compared.

A total of 26 trials were included. At the trial level, four of five (HH, CS, F, and S) and three of four (HH, F, and S) methods for estimating odds ratios and standard errors performed well, respectively. We found considerable differences in the performance of all methods across trials with more accurate estimates for smaller treatment effects. At the level of meta-analysis, three of four methods (CS, F, and S) performed acceptably.

Odds ratios and standard errors can be approximated from continuous end points, but we recommend sensitivity and subgroup analyses to test robustness of the findings.